If you mainly use chatbots and keep worrying you're missing the next big model, this is the part that matters. K3's first product problem is not weakness. It is the default: too much paid thinking for tiny jobs. That is a very ordinary way to waste a beginner's time and money. [C002]
Why should a non-engineer care? Because most people do not start with giant coding tasks. They start with small stuff: summarize this, rewrite that, draw a simple thing. In Simon Willison's pelican drawing test, one tiny prompt reportedly used 95 input tokens, 16,658 output tokens, and 13,241 reasoning tokens, meaning paid hidden thinking, for a total cost of about $0.25. Small test, big clue. [C001]
The important twist is that this is not automatically a "K3 is dumb" story. Moonshot said K3 launched with only max thinking effort, with low and high modes planned later. That shifts the problem from raw intelligence to product routing. Strong brain, wrong first gear.
Kimi K3, and what we can still learn from the pelican benchmark: a model update is not worth tracking because it lists more features. It is worth tracking if it changes your next decision. [C001] Before you trust the hype, run one tiny repeatable task and watch two things: bill and wait time.
Boundary: this is one pelican smoke test plus Moonshot's launch setup, so do not stretch it into "K3 is bad for every simple task." The point is narrower: Kimi K3 may take your first 25 cents before you notice the default is doing too much. If you know someone choosing models by leaderboard alone, share this with them.